See

discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/266629722

A Framework for Building Adaptive Intelligent

Virtual Assistants
Article March 2014
DOI: 10.2316/P.2014.816-018

READS

201

4 authors, including:
Luc Lamontagne

Francois Laviolette

Laval University

Laval University

50 PUBLICATIONS 205 CITATIONS

96 PUBLICATIONS 933 CITATIONS

SEE PROFILE

SEE PROFILE

Richard Khoury
Lakehead University Thunder Bay Campus
39 PUBLICATIONS 63 CITATIONS
SEE PROFILE

All in-text references underlined in blue are linked to publications on ResearchGate,

letting you access and read them immediately.

Available from: Richard Khoury

Retrieved on: 25 May 2016

A FRAMEWORK FOR BUILDING

ADAPTIVE INTELLIGENT VIRTUAL ASSISTANTS

Luc Lamontagne1, Franois Laviolette1, Richard Khoury2, Alexandre Bergeron-Guyard3

Department of Computer Science and Software Engineering, Universit Laval, Qubec, Canada
{Luc.Lamontagne, Francois.Laviolette}@ift.ulaval.ca
2
Department of Software Engineering, Lakehead University, Thunder Bay, Canada
Richard.Khoury@lakeheadu.ca
3
Defence Research & Development Canada, Valcartier, Canada
Alexandre.BergeronGuyard@drdc-rddc.gc.ca

ABSTRACT
This paper describes a framework to support the
construction of intelligent virtual assistants (IVAs). An
IVA agent is a software assistant capable of interacting
with a user to support sense-making tasks, to determine
information needs, to provide relevant information and to
improve its performance based on user feedbacks.
Currently, there is no integrated software environment
available to develop such agents. We are exploring how
we can integrate machine learning and natural language
processing technologies, available as open source
software, to support the construction of intelligent virtual
assistants. The framework relies on a combination of
question answering (Q/A), information extraction (IE) and
user modeling components. In this paper, we present an
overview of the work that is being conducted to build a
prototype of the framework.
KEY WORDS
Intelligent virtual assistant, question answering systems,
information extraction, topic modeling, machine learning,
natural language processing.

1. Introduction
The motivation for this research effort is the development
of an Intelligent Virtual Assistant (IVA) capability for
supporting the intelligence community in their
sensemaking tasks. The design of such systems relies on
intelligent personal assistant technology [1], i.e. software
agents capable of performing tasks with minimum
guidance from its users. Intelligent virtual assistants
combine artificial intelligence technologies to support
users, to organize information and to adapt to changing
situations.
The field of intelligent assistants is currently an
active area of research and development with some
products being much publicized:
CALO and PAL were projects lead by the
Stanford Research Institute (SRI) to investigate
how machine learning can help to design
software agents performing in an office
automation environment. Various modules were
developed during this project for categorization

and filtering of RSS feeds (C2RSS),

recommendations of documents (PrePak), autocompletion of forms (FOAM) and activity
recognition (Warp).
Siri, a spin-off of the CALO project, is a voiceactivated personal assistant commercialized by
Apple to provide factual information about
various information products such as movies,
restaurants, weather reports and concerts.
Google Now is a recent effort by Google to
provide services on their Android operating
system to deliver factual information on a limited
number of topics such as weather reports, flight
information, restaurant reservations and movies.
Mitini and S-Voice are Siri-like systems
developed by Microsoft and Samsung.

All these systems are proprietary and limited

description of the underlying technology can be found in
the literature. They are targeted for the general public,
hence aiming to emphasize ease of use. Their capabilities
are designed to support tasks requiring minimal domain
training such as personal data management and product
recommendation. Interactions with these personal
assistants rely on voice-user interfaces mostly leaving the
initiative to the users.
As part of our research project, we investigate
technologies to design intelligent virtual assistants for
specialized users, i.e. users performing organized tasks on
a recurrent basis. Our main application domain is to assist
sense-making tasks such as those conducted by military
intelligence analysts. Taking information provided by
incoming news reports and military messages, an
intelligence analyst must produce various intelligence
reports. The category of IVA systems targeted in this
project is an interactive system helping an analyst to
retrieve, extract and reuse relevant information from
various textual and semi-structured sources.
Such IVA systems have to cope with the following
particularities:
Task-oriented information search: Tasks such as
the production of intelligence reports follow
some specific procedures and domain models
that can be exploited to determine the type and
nature of information required at each step.

Restricted domain: situations analyzed by

intelligence analysts take place in limited
geographical environments, involving targeted
objects/units and relating to specific topics of
interest. Information about the operational
context can also be exploited to improve the
relevance of system assistance.
Specialized users: analysts have a high level of
training and deep knowledge of their application
domain. Hence they have specific information
needs in specific situations. And they can
provide useful feedbacks on the relevance of
system recommendations and initiatives. More
importantly, they can demonstrate to the system
how to cope with complex situations.

Our approach is to exploit open source technologies

to develop an intelligent software assistant framework
based on natural language processing (NLP) and machine
learning (ML) techniques. And we are validating this
technology on the description of military situations, an
application domain requiring richer background
knowledge than product recommendations or personal
data management. We give an overview of these research
efforts in the next sections.

2.

Context for IVA

This work is conducted in the context of a Defence

Research and Development Canada (DRDC) project
called Intelligence Virtual Analyst Capability (iVAC) [2]
which is an intricate part of the Future Intelligence
Analysis Capability (FIAC) [3]. The iVAC project
investigates and develops ISA technologies towards the
development of a computerized software assistant
supporting the intelligence analysts in sensemaking tasks,
while being ultimately capable of taking on autonomous
analytical tasks in concert with other analysts (virtual or
human).
As part of this research, an identification of iVAC
sub-capabilities was performed, based on literature
reviews and workshops held with experts from the
military, the industry, and academia [4]. This work
allowed identifying question answering and dynamic user
modeling as component of high value for the development
of an iVAC. The IVA project described in this paper is
directly aligned with iVAC priorities and contributes
directly to the development of a system that adapts to the
user's context, and answers questions based on context
and available information.

3.

Components of the IVA framework

The choice of components to include in the IVA

framework was guided by different motivations. We
present in the next paragraphs, a detailed description of
the components we selected to be part of the framework.

3.1 Integrated framework for the design of IVAs

As illustrated in Figure 1, an IVA aims to support a user
in its domain specific tasks. For our application domain,
an IVA helps an intelligence specialist to analyze a
situation
involving
different
aspects
including
geopolitical, socioeconomic and demographic factors.
As most of the information sources available in our
application domain are textual in nature, the IVA
framework includes a question-answering component that
can be queried by the user to gain information about some
of these factors. Typical queries would pertain to:
Characteristics of domain entities such a military
units,
ethnic
groups,
political
actors,
infrastructures and important dates.
Relations such as roles of leaders, population
estimates of some regions, nature of conflicts
between factions, electricity or natural resources
providers, etc.
Events depicting actions taken by various actors
present in the region of interest.
To facilitate the search for relevant information
passages, an information extraction component processes
domain-related documents (ex. new reports and military
messages) and annotates them with domain specific tags.
We expect that the insertion of annotation scheme will
improve the relevance of system responses. Categories of
annotations targeted in the project are named entities and
semantic relations. Some annotations are generic (ex.
named entities) while some specific annotation sets could
result from information needs of the user.
As the technology is targeted for specialized users,
we want to personalize the recommendations of the IVA
system by constructing user models from unsupervised
machine learning techniques. Construction of the user
models rely on the analysis of traces depicting the
information and documents manipulated by the user.
Activity traces contain a combination of the following
information:
-
Documents received to the user from external
sources. In our current experimentations, most of the
external sources are RSS feeds that are managed
using RSSOwl as a news feed reader;
-
Documents consulted by the user from local
collections or external sources;
-
Queries submitted by the user to the IVA system;
-
Inputs provided by the user in the textual fields of
intelligence templates (i.e. some reports produced by
the analyst).
User models are the pivot elements of the IVA
framework as they capture the potential information needs
by the user, support the analysis and the disambiguation
of the questions submitted to the Q/A component, and
provide guidance for the annotation of documents in the
local collections. User models can also be augmented
through explicit interactions with users.

Figure 1. Main components of the IVA framework
The three components of the framework are further
described in the next sections. Moreover, we then explain
some of the interactions taking place among these
components.
3.2

Automatic annotation of textual documents

Searching for answers in textual documents can be

facilitated if the type of information being sought is
clearly highlighted. For example, one might be searching
for some specific entities (ex. individuals, organizations,
locations) or some relationships between individuals and
events. By performing a priori annotation of documents
as a batch process, one would expect to increase the
relevance of the passages returned by the Q/A system
without impacting on response time.
The text annotation component of the IVA
framework includes two types of information extraction
systems. First a named entity recognition (NER) system
[5] is trained to detect the domain entities of interest to the
user. Our implementation relies on the Stanford NER
libraries [6] that can be trained with tagged data sets. NER
technology is mature [7] and provides satisfying results if
combined with domain-specific resources (ex. gazetteers
and name lists) and representative training data.
Second, we extract semantic relations from texts. It
consists of finding triplets of the form x relation y that
are recurrent structures in textual documents. For
instance, the sentence Barrack Obama is the president of
the United States would be decomposed as a triplet
containing the two arguments Barrack Obama and
United States and the relation be the president of.
Significant progress has been made recently by open
information extraction systems on the extraction of

meaningful relations out of documents [8, 9, 10, 11].

Some algorithms are showing good results for identifying
relation phrases on a large-scale corpus without any
specific domain training other than using simple linguistic
patterns. Recent algorithms extend relations extracted by
these systems to phrases mediated by verb and noun
patterns. Moreover, these systems offer the advantage of
not requiring any pre-specified vocabulary or set of
relations. Hence they can be used for the discovery of
important relations to be subsequently validated by
domain experts.
If more accurate extraction capabilities are required,
the extraction process could be augmented by semantic
role labelling algorithms [12] that associate verbs to a
semantic structure and map the other phrases of the
sentence to roles in this structure. Supervised machine
learning techniques (such as SVM or MEMM classifiers)
can be used to tag text documents with semantic classes
and roles relevant to the application [13, 14]. However we
do not pursue this approach at the moment due to the
good results obtained with the information extraction
components (see Section 4 of this paper).
3.3 Learning of User Models
Learning a user model amounts to capture the potential
information needs of the users. For instance, an IVA for
intelligence operations should have some representation
of the military, geopolitical and economical factors that
might be required at various stages of operations. An IVA
system must be able to learn these models from users
feedbacks, together with past questions and prior
manipulations of documents. Such models can be
exploited to improve systems performance.

In our current framework implementation, user

models are constructed using topic-modeling algorithms
[15]. Topic modeling tools are unsupervised learning
algorithms that represent topics as probability
distributions over various clusters of words (the topics).
We make use of the MALLET library that includes
algorithms such as Latent Dirichlet Allocation (LDA),
hierarchical LDA and Pachinko allocation [16, 17]. Topic
models are constructed from the activity traces provided
to the framework by the RSS news feed reader.
Taking advantage of the information extractors of the
framework, we include in the topic models extracted roles
such as named entity classes and relation phrases. These
augmented topics are useful as they can capture high-level
interests of the users. For instance, a model could contain
individual terms such as negotiation and talk, roles
such as ORGANISATION and relations such as be
scheduled.
While this has not been undertaken yet, we would
like to model the context in which the user needs specific
information. To acquire these context descriptions,
monitoring procedures have to be developed for acquiring
demonstration traces from users. In learning by
demonstration paradigm [18], a system learns how to
behave from examples provided by human teachers. The
demonstration consists of examples provided as
sequences of state-action pairs, indicating to the system
how a task should be accomplished [18,19,20,21]. For
instance, in our application, demonstrations depict the
information targeted by intelligence analysts at the
various stages of their analysis and the textual passages
inserted in the intelligence reports. From the
demonstration traces, algorithms such as those used for
on-line topic detection and tracking [22] could be used to
capture context episodes and to reuse them as guidelines
when facing new situations.
3.4 Adaptive Question Answering
As a virtual assistant is mainly an information service, the
main functionalities of the IVA framework are provided
by a question answering (Q/A) system. The modules
encountered in the pipeline of a Q/A system [23] are
implementing three types of functions: question analysis,
passage retrieval and answer selection.
Question analysis consists of understanding what is
being asked by the user. At a minimum, the system needs
to classify the topic of the question and to extract
important keywords featured in the question. Additional
steps are also added to determine the expected answer
type such as factoid information, opinions or short
definitions. It also provides information on the kind of
information requested (ex. a person, a location, a yes/no
answer). Accurately pinpointing the characteristics of
answers can greatly improve the accuracy of a Q/A
system, notably by making possible to use type-specific or
kind-specific information retrieval algorithms and by
filtering out answers of the wrong type [24].

Passage retrieval retrieves and filters out text

fragments that might contain the answer sought by the
user. A typical Q/A system must include an information
retrieval (IR) subsystem to search for documents relevant
to the question asked, as well as an information extraction
(IE) subsystem to pinpoint relevant passages within
longer texts. In some Q/A systems, these two steps can be
combined. For example, we can break up a long document
into parts (such as individual paragraphs) and allow the IR
subsystem to retrieve these individual parts. While
creating a simpler architecture, this allows the system to
easily account for long documents discussing multiple
different topics in turn. However, we kept these as two
different steps in our Q/A system. Many different IR
algorithms have been proposed in the literature to handle
the different information sources. They can range from
simple keyword searches of plain text documents to
collaborative filtering of online resources. Likewise, the
IE step will vary depending on the knowledge base being
used. Simpler systems can operate at the word level, for
example by returning fields with the correct keyword
label in semi-structured documents or sentences featuring
named entities found in the question. More sophisticated
systems often operate at the sentence-syntax level, aiming
to discover text fragments that do not simply mention the
correct keywords of the question but mention them with
the sentence structure of the answer that is expected given
the structure of the corresponding question [25].
Answer generation presents the answers in a format
acceptable to the user. The passage retrieval phase returns
several relevant text fragments. In its simplest form, a
Q/A system ranks these fragments and returns the ranked
list to the user. Several improvements are applied on this
basic strategy, such as grouping together duplicate
answers from different text fragments, or presenting text
fragments with their surrounding sentences to capture the
context. More advanced systems will actually rephrase the
information discovered into an original answer statement.
Some systems use semantic templates that specify which
answer component to use in a given field [26], while
others prefer syntactic templates that specify word order
in the answer [27]. In addition, some Q/A systems
implement different templates for the different devices.
For example, a verbose template style would be used
when displaying results on a desktop computer, while a
concise template that omits unnecessary contextual
information would be used when displaying the result on
the smaller screen of a mobile device [26].
Building on this three-step pipeline architecture, an
adaptive question answering system [28] is a Q/A system
that can improve its performance based on the operational
context. Q/A systems can be made adaptive by injecting
models describing the goals and preference of the
questioner and the characteristics of the task being
performed. Adaptive Q/A has not been the subject of
extensive research, as the recent trend has been to have
stateless systems where users ask only one question per
session [28]. This is a trend justified by studies showing
that this is the way a majority of users behave, especially

when it comes to using online query systems [29].

However, this differs in the application domain targeted
by our project. The same specialized users query the Q/A
component of the IVA framework repeatedly for
consistent purposes. Such system can clearly benefit from
adapting its behaviour to its users and purposes.
The adaptive Q/A component of the IVA framework
is built on Open Ephyra, an open source Q/A
infrastructure. Currently, information provided to the
system is taken from Wikipedia, the CIA World Factbook
and also includes all the documents received from
subscribed RSS news feed. Lucene and Indri search
systems are responsible to index the RSS news reports in
the local collections and to support the passage retrieval
phase of the Q/A system. More specifically, full
documents are first searched using Lucene. Passages are
then extracted using both Lucene and Indri snippet
extraction capabilities. This choice is mainly due to the
complementary nature of the passages extracted by both
systems and to the greater stability of Lucene for indexing
large collections of documents.
The IVA framework extends the Q/A infrastructure
with techniques exploiting the user models acquired from
users activity traces (these techniques are described in
Section 5). Our aim is to go beyond answering explicit
factoid questions, and to return information that is
relevant to the user asking the question within the context
of the user model. Techniques combining information
from the user model, from the systems internal annotated
knowledge base and from external knowledge bases (such
as Wikipedia) can be used to improve the performance of
the three steps of the Q/A process.

4. Preliminary experiments
We conducted some preliminary experiments early in the
project to determine the baseline performance of the IVA
framework for our military intelligence application
domain.
The annotation component relies on two extraction
technologies. We first evaluated the NER extraction
component that we have selected for the project (Stanford
NER). We obtained accuracy greater than 80% on a
corpus of geopolitical news reports representative of our
military application domain. We estimate that these
results fully meet our expectations. However additional
training could be performed on our domain corpus if
higher levels of accuracy are required for some other
application domains.
We also evaluated the performance of REVERB [6]
as relation extractor for our framework. Experimentations
conducted on our corpus of geopolitical news reports
resulted in an accuracy of approximately 60% of the
relations present in these documents. Moreover, most of
the relations extracted by the system were valid and could
be made to contribution in the framework. This result
suggests that most of the important relations of an
application domain can be identified and exploited as part
of the Q/A process.

The performance of the Q/A component was

evaluated on different sources of information. To do so,
we built a list of 200 questions that were submitted to the
system. Most of the questions were related to various
geopolitical conflicts occurring during the months our
experiments were performed (ex. Syria and Mali). We
made sure that all the answers were present, in some
forms, in the corpus or resources exploited by the
information sources. Hence our goal was to assess the
capability of the Q/A to locate, circumscribe and rank the
relevant text passages knowing that these could be
provided by the information sources.
We estimated from our experimentations that,
overall, less than 30% of the questions were correctly
answered by the most highly ranked passage of the Q/A
system. The fragmented results for each information
source are presented in Table 1.
Questions that were considered to present either low
or average level of difficulty (ex. simple factual
information and definitions) were found in a reasonable
proportion by the system (between 30-40%). However,
none of the challenging questions considered as difficult
or very difficult could be answered by the system.
Table 1
Preliminary Q/A results for different sources of information
Wikipedia & CAI World Factbook
35%
Domain corpus with Indri search engine
25%
Domain corpus with Lucene search engine
22%
Domain corpus with Indri and Lucene search engine
24%

However, following a detailed analysis of the Q/A

results, it appears that these baseline results can greatly be
improved since more that 50% of the answers were
contained in the first 5 passages returned by the system.
And over 90% of the answers were present in the tens of
passages extracted by the system. Hence providing better
answer selection schemes could significant improve the
accuracy of the system. We explore in the next section
some research avenues we are currently pursuing to
improve these results.

5. Adaptation of IVA behaviour

Following the preliminary experiments that we conducted
on our military intelligence application domain, we
identified candidate techniques in order to be added to
help adapting the IVA framework to new users and new
application domains. Three of them, taking advantage of
synergy between the components in the framework, are
described in the next paragraphs.
5.1 Questions expansion and paraphrasing
A common technique to improve the performance of Q/A
and IR systems is query expansion, which consists of
supplying common domain-specific keywords missing in
the question. The need for this step stems from the fact
that most questions are very short, between one and four

keywords long on average [30] (after removal of the

words bearing no specific meanings). Moreover, these
keywords can have ambiguous meanings [31]. These two
factors limit the accuracy of any information retrieval
operation based on short question. Query expansion can
benefit to the system by allowing it to retrieve documents
and answers that would otherwise be missed because they
do not have any keywords in common with the original
question [32]. More specifically, we design the question
analysis phase of the Q/A system in such a way that it
includes a step where the question is rewritten into
multiple paraphrased forms. This step allows the system
to recognize different forms of the same question and then
to expand them all so to improve the recall of text
passages by the information retrieval system.
In our current work, we are exploring how questions
can be expanded based on the paraphrasing of their main
relation. For instance, we would like to expand the
question When was Louisiana acquired by the USA into
multiple paraphrased versions using relations such as
bought by, purchased by and became part of. To
do so, we need to create a linguistic resource that provides
clusters of relations based on synonyms, called synsets.
As such a resource is not currently available, we have
built a preliminary version using relations extracted from
Wikipedia with the text annotation component of the
framework. Given a list of relations and the context in
which they appear, we can statistically evaluate the
semantic relatedness of the relations. The context of a
relation is built from the list of words surrounding its
occurrences in sentences. And the semantic similarity of
two context vectors can be estimated using similarity
measures (we chose a normalized cosine function). For
instance, the synsets obtained with this scheme for the
relation announce ceasefire is:
[announce victory (0,316) declare neutrality (0,258)
announce settlement (0,223) declare ceasefire (0,223), ]

As tens of thousands relations can be mutually

compared for a single application domain, this process
can lead to a computational explosion. We reduce the
complexity of the learning scheme by using Wordnet to
limit the comparisons to relations containing related
words. A simplified version of the pseudo-code for the
overall approach is provided in Figure 2.
function BUILD-RELATION-SYNSETS(relations, triplets)
for every relation r1 in relations
merge WordNet synsets for all the words in r1 as syn1 .
for every other relation r2 in relations
retain as candidates1 the relations r2 containing words
from syn1.
for every relation r1 in relations
for every other relation r2 in candidates1
build context vectors v1 and v2 from triplets of r1 and r2 .
compute similarity s between v1 and v2 .
retain the top relations relation_synset1 ranked by
decreasing similarity s.

Figure 2. Pseudo-code for building relation synsets

5.2 Adaptive filtering and scoring of the answers

As mentioned previously, we estimated from our
experimentations that less than 30% of the questions were
correctly answered by the most highly ranked passage of
the Q/A system. However this could greatly be improved
since more that 50% of the answers were contained in the
first 5 passages returned by the system. Hence providing
better answer selection schemes could significant improve
the accuracy of the system.
Answer selection consists of applying a series of
filters to rate passages based on the textual content and to
remove those that do not match the characteristics of the
questions. Hence, the quality of the results highly depends
on the choice of filters included in the pipeline and on the
order in which they are placed.
To address this problematic, we have implemented a
genetic optimization algorithm that seeks for the optimal
configuration of filters to be included in the system. A
genetic algorithm [33] is a search technique that uses
specific heuristics, inspired from the process of natural
selection, to determine how to conduct its optimisation.
The optimization routine consists in generating an initial
population of individuals (representing potential solutions
to our problem) and making this population evolve over
time by modifying the characteristics of the individuals.
function SELECT-FILTERS (filters, questions, answers)
1. Generate an initial array of populations containing different
combinations of filters;
2. Evaluate the fitness of each combination using the training
pairs of questions and answers;
3. Select at random (using a Russian roulette function) two
combinations for each population;
4. Swap parts of the selected combinations using a one-point
crossover function and add them to the population;
5. Mutate some filters in the newly generated combinations;
6. Migrate some combinations from one population to the other
7. Select a number of combinations to remove (equal to the
number of new combinations);
8. Repeat from step 2 until some stopping criterion is satisfied;
9. Return the combination of filters having the best fitness
value in the final populations

Figure 3. Pseudo-code for the selection of filters

In our application, each individual corresponds to a
sequence of a subset of the filters. The characteristics of
each individual indicate the relative importance of a
specific filter. To conduct the optimization, a routine has
also been implemented to evaluate the fitness of a
combination of filters on a set of training questions and
answers. For Q/A, the fitness of combination of filters
corresponds to the accuracy of the answers selected by the
filter combination. Hence given a list of question-answer
pairs provided by the user, it is possible to customize the
answer selection phase of the Q/A system to the specific
usage of the system by some users. The pseudo-code of
the optimization routine is described in Figure 3.

build user models that are not strictly constructed from

topic modeling algorithms.

5.3 User-specific ranking of the answers

To exploit user models as part of the question answering
process, we also included a probabilistic scoring approach
based on the topics for the questions and the candidate
answers. First, the set of relevant topics to the user is
learned based on a topic modeling approach such as
Latent Dirichlet Allocation (LDA). The models are
learned from the news reports being read as well as from
the questions being asked by the user. Using these
models, it is possible to estimate the similarity of a new
question asked by the user to the candidate answers
returned by the system.
The score of a candidate answer, i.e., a text fragment
returned by the passage retrieval phase of the Q/A system,
can be estimated as the highest probable topic given some
answer and some question. More formally, the topicmodel score of a candidate answer can be evaluated as:
!! ! |,
(1)
where ta is the answer topic, q is the question formulated
by the user and a is a candidate answer returned by the
search engine.
Equation (1) can be further decomposed, using Bayes
and chained rules, as:
!! ! |, = !! |! |! !

(2)

which corresponds to the multiplication of three

probability values designating:
a) The relatedness of the query to the topic,
b) The relatedness of the answer to the topic, and
c) The likelihood of occurrence of the topic in the
Q/A process.
All three probability distributions can easily be
obtained from a topic-modeling algorithm applied to the
information exchanged between the IVA system and the
user. We are currently performing an evaluation of this
answer-filtering scheme.

6. Conclusion
This paper gave an overview of the efforts we are
devoting to the development of a framework for building
intelligent virtual assistants. We described the three main
components of it and explained how they are in the
framework. Moreover, we discussed some of the practical
issues we are facing. We have discussed three adaptation
techniques that are being added to the framework in order
to improve its performance. Details were provided on
technical decision made during the implementation.
Future work on this project will include a global
evaluation of the IVA system including the adaptation
scheme presented in Section 5 of the paper. And further
experimentations will also be required to explore how to

Acknowledgement(s)
We would like to thank the Natural Sciences and
Engineering Research Council of Canada (NSERC) and
the Canadian Department of National Defence for the
financial support of the work through their joint DNDNSERC research program.

References
[1] K. Myers, P. Berry, J. Blythe, K. Conley, M.
Gervasio, D. McGuinness, D. Morley, A. Pfeffer, M.
Pollack, & M. Tambe, An Intelligent Personal Assistant
for Task and Time Management, AI Magazine, 28(2),
2007, 47-61.
[2] D. Gouin, V. Lavigne, & A. Bergeron-Guyard,
Human-computer interaction with an intelligence virtual
analyst, KSCO, 2012.
[3] D. Poussart, Future intelligence analysis capability towards a cohesive R&D program definition, Technical
Report, TM 2012-9999, DRDC Valcartier, 2012.
[4] G. Michaud, A. Burkov, M. Brub, L. Lamontagne,
& F. Laviolette, Identification of iVAC capabilities,
Technical report, CR 2011-243, DRDC Valcartier, 2012.
[5] D. Nadeau & S. Sekine, A survey of named entity
recognition
and
classification,
Linguisticae
Investigationes, 30(1), 2007, 3-26.
[6] J. R. Finkel, T. Grenager & C. Manning,
Incorporating Non-local Information into Information
Extraction Systems by Gibbs Sampling, Proceedings of
the 43nd Annual Meeting of the Association for
Computational Linguistics, 2005, 363-370.
[7] M. Marrero, S. Sanchez-Cuadrado, J. Morato & Y.
Andreadakis, Evaluation of Named Entity Extraction
Systems, Research in Computer Science, vol. 41, 2009,
47-58.
[8] M. Banko, M. Cafarella, S. Soderland, M.
Broadhead, & O. Etzioni, Open information extraction
from the Web, Proceedings of IJCAI, 2009, 2670-2676.
[9] A. Fader, S. Soderland, & O. Etzioni, Identifying
relations for open information extraction, Proceedings of
EMNLP, 2011, 1535-1545.
[10] F. Wu & D. S. Weld. Open information extraction
using Wikipedia. Proceedings of the ACL, 2010, 118-127.
[11] M. Mausam, R. Schmitz, R. Bart, S. Soderland, & O.
Etzioni, Open language learning for information
extraction, Proceedings of the 2012 Joint Conference on
Empirical Methods in Natural Language Processing and
Computational Natural Language Learning, 2012, 523534.
[12] D. Gildea & D. Jurafsky, Automatic Labeling of
Semantic Roles, Computational Linguistics, 28(3), 2002,
245-288.
[13] T. Mitsumori, M. Murata, Y. Fukuda, K. Doi & H.
Doi, Semantic Role Labeling Using Support Vector

Machines, Proceedings of the Ninth Conference on

Computational Natural Language Learning, 2005, 197200.
[14] S. Weiwei, H. Li & Z. Sui, The integration of
dependency relation classification and semantic role
labeling using bilayer maximum entropy Markov models,
Proceedings of the Twelfth Conference on Computational
Natural Language Learning, 2008, 243-247.
[15] D.M. Blei, Introduction to Probabilistic Topic
Models, Communications of the ACM, 55(4), 2012, 77
84.
[16] D.M. Blei, A.Y. Ng, M.I. Jordan & J.
Lafferty, Latent Dirichlet allocation, Journal of Machine
Learning Research, vol. 3, 2003, 9931022.
[17] W. Li & A. McCallum, Pachinko Allocation: DAGStructured Mixture Models of Topic Correlations,
Proceedings of the 23rd International Conference on
Machine Learning, 2006, 577-584.
[18] H. Lieberman, Your Wish is My Command:
Programming by Example (San Francisco, CA: Morgan
Kaufmann, 2001).
[19] M. Gervasio, T. Lee & S. Eker, Learning email
procedures for the desktop, Proceedings of the AAAI
Workshop on Enhanced Messaging, AAAI Press, 2008,
22-27.
[20] S. Eker, T. J. Lee & M. Gervasio, Iteration Learning
by Demonstration, Proc. of the AAAI 2009 Spring
Symposium on Agents that Learn from Human Teachers,
2009.
[21] K.L. Myers, J. Kolojejchick, C. Angiolillo, T.
Cummings, T. Garvey, M.T. Gervasio, W. Haines, C.
Jones, J. Knittel, D.N. Morley, W. Ommert, & S. Potter.
Learning by Demonstration Technology for Military
Planning and Decision Making: A Deployment Story,
Proceedings of IAAI, 2011, 1597-1604.
[22] L. AlSumait, D. Barbara & C. Domeniconi, On-line
LDA: Adaptive Topic Models for Mining Text Streams
with Applications to Topic Detection and Tracking,
Proceedings of the 2008 Eight IEEE International
Conference on Data Mining, 2008, 3-12.

[23] L. Hirschman & R. Gaizauskas, Natural Language

Question Answering. The View from Here, Natural
Language Engineering, 7(4), 2001, 275-300.
[24] M. Razmara, A. Fee & L. Kosseim, Concordia
University at the TREC 2007 QA track; Proceedings of
the Sixteenth Text Retrieval Conference, 2007.
[25] S. Harabagiu, D. Moldovan, M. Pasca, R. Mihalcea,
M. Surdeanu, R. Bunescu, R. Grji, V. Rus & P.
Morarescu, FALCON: Boosting knowledge for answer
engines, Proceedings 9th Text Retrieval Conference
(TREC-9), 2000, pp. 479-488.
[26] B. Magnini, M. Speranza & V. Kumar, Towards
Interactive Question Answering: An Ontology-Based
Approach, IEEE International Conference on Semantic
Computing (ICSC '09), 2009, 612-617.
[27] S. Song & Y.-N. Cheah, An unsupervised center
sentence-based clustering approach for rule-based
question answering, IEEE Symposium on Computers &
Informatics (ISCI), 2011, 125-129.
[28] S. Quarteroni & S. Manandhar, User modelling for
adaptive question answering and information retrieval,
Proceedings of FLAIRS06, 2006, 199-202.
[29] B. J. Jansen, A. Spink, & T. Saracevic, Real life, real
users, and real needs: a study and analysis of user queries
on the web, Information Processing and Management,
36(2), 2000, 207-227.
[30] D. Shen, R. Pan, J.T. Sun, J.J. Pan, K. Wu, J. Yin, &
Q. Yang, Q2C@UST: our winning solution to query
classification in KDDCUP 2005, ACM SIGKDD, 7(2),
2005, 100-110.
[31] F.B. Dian Paskalis, & M.L. Khodra, Word sense
disambiguation in information retrieval using query
expansion, International Conference on Electrical
Engineering and Informatics, 2011, 1-6.
[32] J. Luo, B. Meng, & X. Tu, Selecting good expansion
terms based on similarity kernel function, 7th
International Conference on Natural Language
Processing and Knowledge Engineering (NLP-KE), 2011,
156-160.
[33] A. Eiben & J. Smith, Introduction to Evolutionary
Computing (Springer, 2003).